The small one. The fast one.
Anima is a 2B-parameter Cosmos-Predict2 MMDiT distilled for few-step generation. At ~5.4 GB it is the lightest model TheNoise ships with, and at ~5 seconds per 1024×768 image on Strix Halo it is also the fastest. If you want to iterate on prompts quickly — dozens of takes, minimal wait — this is the model to start on.
- Text-to-image only (no editing)
- Built-in defaults: 8 steps, guidance 1.0 (CFG off) — no tuning needed
- int8-convrot variant available (smaller download, lower memory)
Commands below assume a dev checkout (
./thenoise.sh,.venv/bin/python). On a portable bundle use./bin/thenoiseand./bin/python3instead.
| Spec | Data |
|---|---|
| Architecture | 2B Cosmos-Predict2 MMDiT (flow matching) |
| Download size | ~5.4 GB (bf16), less with --int8-convrot |
| VAE | Qwen-Image VAE |
| Text encoder | Qwen3-0.6B |
| Editing | — |
| KV cache | — |
| Default settings | 1024×1024, 8 steps, guidance 1.0, flow shift 3.0 |
| Variant | Steps | Character |
|---|---|---|
turbo-v1.0 (default) |
8 | fewest steps, the workhorse |
aesthetic-v1.1 |
more | aesthetically tuned |
base-v1.0 |
~20, CFG ~4 | base (non-distilled) checkpoint |
Ready-made Lemonade recipes for this model - each pins the checkpoints and the generation defaults:
- Anima-Turbo.json - Turbo BF16, 8 steps
- Anima-Turbo-INT8.json - Turbo INT8-ConvRot, 8 steps
- Anima-Aesthetic.json - Aesthetic BF16, 20 steps, CFG 4
- Anima-Aesthetic-INT8.json - Aesthetic INT8-ConvRot, 20 steps, CFG 4
- Anima-Base.json - Base BF16, 20 steps, CFG 4
- Anima-Base-INT8.json - Base INT8-ConvRot, 20 steps, CFG 4
- Turbo @ 8 steps: ~5 s (1024×768)
- Base @ 20 steps, CFG 4: ~20 s
Generated — (Turbo, 8 steps):
Generated — (Base, 20 steps, CFG 4):
.venv/bin/python scripts/download.py --model animaOther variants / quantization:
# a specific DiT variant (the value becomes part of the --dit filename)
.venv/bin/python scripts/download.py --model anima --variant aesthetic-v1.1
# int8-convrot DiT instead of bf16
.venv/bin/python scripts/download.py --model anima --int8-convrotGenerate:
./thenoise.sh generate \
--dit ./models/anima/split_files/diffusion_models/anima-turbo-v1.0.safetensors \
--vae ./models/anima/split_files/vae/qwen_image_vae.safetensors \
--text-encoder ./models/anima/split_files/text_encoders/qwen_3_06b_base.safetensors \
--prompt "a fox walking in the snow" \
--out fox.pngServe over HTTP with the web UI (open http://localhost:8000/):
./thenoise.sh serve \
--dit ./models/anima/split_files/diffusion_models/anima-turbo-v1.0.safetensors \
--vae ./models/anima/split_files/vae/qwen_image_vae.safetensors \
--text-encoder ./models/anima/split_files/text_encoders/qwen_3_06b_base.safetensors \
--host 127.0.0.1 --port 8000For every flag (size, steps, seed, LoRAs, upscaling, post-processing), see the CLI reference and the HTTP API reference.